1 //===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation implements the well known scalar replacement of
11 // aggregates transformation. This xform breaks up alloca instructions of
12 // aggregate type (structure or array) into individual alloca instructions for
13 // each member (if possible). Then, if possible, it transforms the individual
14 // alloca instructions into nice clean scalar SSA form.
16 // This combines a simple SRoA algorithm with the Mem2Reg algorithm because
17 // often interact, especially for C++ programs. As such, iterating between
18 // SRoA, then Mem2Reg until we run out of things to promote works well.
20 //===----------------------------------------------------------------------===//
22 #define DEBUG_TYPE "scalarrepl"
23 #include "llvm/Transforms/Scalar.h"
24 #include "llvm/Constants.h"
25 #include "llvm/DerivedTypes.h"
26 #include "llvm/Function.h"
27 #include "llvm/GlobalVariable.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/IntrinsicInst.h"
30 #include "llvm/Pass.h"
31 #include "llvm/Analysis/Dominators.h"
32 #include "llvm/Target/TargetData.h"
33 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
34 #include "llvm/Transforms/Utils/Local.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/GetElementPtrTypeIterator.h"
37 #include "llvm/Support/IRBuilder.h"
38 #include "llvm/Support/MathExtras.h"
39 #include "llvm/Support/Compiler.h"
40 #include "llvm/ADT/SmallVector.h"
41 #include "llvm/ADT/Statistic.h"
42 #include "llvm/ADT/StringExtras.h"
45 STATISTIC(NumReplaced, "Number of allocas broken up");
46 STATISTIC(NumPromoted, "Number of allocas promoted");
47 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
48 STATISTIC(NumGlobals, "Number of allocas copied from constant global");
51 struct VISIBILITY_HIDDEN SROA : public FunctionPass {
52 static char ID; // Pass identification, replacement for typeid
53 explicit SROA(signed T = -1) : FunctionPass(&ID) {
60 bool runOnFunction(Function &F);
62 bool performScalarRepl(Function &F);
63 bool performPromotion(Function &F);
65 // getAnalysisUsage - This pass does not require any passes, but we know it
66 // will not alter the CFG, so say so.
67 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
68 AU.addRequired<DominatorTree>();
69 AU.addRequired<DominanceFrontier>();
70 AU.addRequired<TargetData>();
77 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
78 /// information about the uses. All these fields are initialized to false
79 /// and set to true when something is learned.
81 /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
84 /// needsCleanup - This is set to true if there is some use of the alloca
85 /// that requires cleanup.
86 bool needsCleanup : 1;
88 /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
91 /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
95 : isUnsafe(false), needsCleanup(false),
96 isMemCpySrc(false), isMemCpyDst(false) {}
101 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; }
103 int isSafeAllocaToScalarRepl(AllocationInst *AI);
105 void isSafeUseOfAllocation(Instruction *User, AllocationInst *AI,
107 void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI,
109 void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI,
110 unsigned OpNo, AllocaInfo &Info);
111 void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI,
114 void DoScalarReplacement(AllocationInst *AI,
115 std::vector<AllocationInst*> &WorkList);
116 void CleanupGEP(GetElementPtrInst *GEP);
117 void CleanupAllocaUsers(AllocationInst *AI);
118 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base);
120 void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI,
121 SmallVector<AllocaInst*, 32> &NewElts);
123 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst,
125 SmallVector<AllocaInst*, 32> &NewElts);
126 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocationInst *AI,
127 SmallVector<AllocaInst*, 32> &NewElts);
128 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI,
129 SmallVector<AllocaInst*, 32> &NewElts);
131 bool CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy,
132 bool &SawVec, uint64_t Offset, unsigned AllocaSize);
133 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset);
134 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType,
135 uint64_t Offset, IRBuilder<> &Builder);
136 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
137 uint64_t Offset, IRBuilder<> &Builder);
138 static Instruction *isOnlyCopiedFromConstantGlobal(AllocationInst *AI);
143 static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates");
145 // Public interface to the ScalarReplAggregates pass
146 FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) {
147 return new SROA(Threshold);
151 bool SROA::runOnFunction(Function &F) {
152 TD = &getAnalysis<TargetData>();
154 bool Changed = performPromotion(F);
156 bool LocalChange = performScalarRepl(F);
157 if (!LocalChange) break; // No need to repromote if no scalarrepl
159 LocalChange = performPromotion(F);
160 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
167 bool SROA::performPromotion(Function &F) {
168 std::vector<AllocaInst*> Allocas;
169 DominatorTree &DT = getAnalysis<DominatorTree>();
170 DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
172 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
174 bool Changed = false;
179 // Find allocas that are safe to promote, by looking at all instructions in
181 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
182 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
183 if (isAllocaPromotable(AI))
184 Allocas.push_back(AI);
186 if (Allocas.empty()) break;
188 PromoteMemToReg(Allocas, DT, DF);
189 NumPromoted += Allocas.size();
196 /// getNumSAElements - Return the number of elements in the specific struct or
198 static uint64_t getNumSAElements(const Type *T) {
199 if (const StructType *ST = dyn_cast<StructType>(T))
200 return ST->getNumElements();
201 return cast<ArrayType>(T)->getNumElements();
204 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
205 // which runs on all of the malloc/alloca instructions in the function, removing
206 // them if they are only used by getelementptr instructions.
208 bool SROA::performScalarRepl(Function &F) {
209 std::vector<AllocationInst*> WorkList;
211 // Scan the entry basic block, adding any alloca's and mallocs to the worklist
212 BasicBlock &BB = F.getEntryBlock();
213 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
214 if (AllocationInst *A = dyn_cast<AllocationInst>(I))
215 WorkList.push_back(A);
217 // Process the worklist
218 bool Changed = false;
219 while (!WorkList.empty()) {
220 AllocationInst *AI = WorkList.back();
223 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
224 // with unused elements.
225 if (AI->use_empty()) {
226 AI->eraseFromParent();
230 // If this alloca is impossible for us to promote, reject it early.
231 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
234 // Check to see if this allocation is only modified by a memcpy/memmove from
235 // a constant global. If this is the case, we can change all users to use
236 // the constant global instead. This is commonly produced by the CFE by
237 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
238 // is only subsequently read.
239 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) {
240 DOUT << "Found alloca equal to global: " << *AI;
241 DOUT << " memcpy = " << *TheCopy;
242 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2));
243 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
244 TheCopy->eraseFromParent(); // Don't mutate the global.
245 AI->eraseFromParent();
251 // Check to see if we can perform the core SROA transformation. We cannot
252 // transform the allocation instruction if it is an array allocation
253 // (allocations OF arrays are ok though), and an allocation of a scalar
254 // value cannot be decomposed at all.
255 uint64_t AllocaSize = TD->getTypePaddedSize(AI->getAllocatedType());
257 // Do not promote any struct whose size is too big.
258 if (AllocaSize < SRThreshold)
261 if ((isa<StructType>(AI->getAllocatedType()) ||
262 isa<ArrayType>(AI->getAllocatedType())) &&
263 // Do not promote any struct into more than "32" separate vars.
264 getNumSAElements(AI->getAllocatedType()) < SRThreshold/4) {
265 // Check that all of the users of the allocation are capable of being
267 switch (isSafeAllocaToScalarRepl(AI)) {
268 default: assert(0 && "Unexpected value!");
269 case 0: // Not safe to scalar replace.
271 case 1: // Safe, but requires cleanup/canonicalizations first
272 CleanupAllocaUsers(AI);
274 case 3: // Safe to scalar replace.
275 DoScalarReplacement(AI, WorkList);
281 // If we can turn this aggregate value (potentially with casts) into a
282 // simple scalar value that can be mem2reg'd into a register value.
283 // IsNotTrivial tracks whether this is something that mem2reg could have
284 // promoted itself. If so, we don't want to transform it needlessly. Note
285 // that we can't just check based on the type: the alloca may be of an i32
286 // but that has pointer arithmetic to set byte 3 of it or something.
287 bool IsNotTrivial = false;
288 const Type *VectorTy = 0;
289 bool HadAVector = false;
290 if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector,
291 0, unsigned(AllocaSize)) && IsNotTrivial) {
293 // If we were able to find a vector type that can handle this with
294 // insert/extract elements, and if there was at least one use that had
295 // a vector type, promote this to a vector. We don't want to promote
296 // random stuff that doesn't use vectors (e.g. <9 x double>) because then
297 // we just get a lot of insert/extracts. If at least one vector is
298 // involved, then we probably really do have a union of vector/array.
299 if (VectorTy && isa<VectorType>(VectorTy) && HadAVector) {
300 DOUT << "CONVERT TO VECTOR: " << *AI << " TYPE = " << *VectorTy <<"\n";
302 // Create and insert the vector alloca.
303 NewAI = new AllocaInst(VectorTy, 0, "", AI->getParent()->begin());
304 ConvertUsesToScalar(AI, NewAI, 0);
306 DOUT << "CONVERT TO SCALAR INTEGER: " << *AI << "\n";
308 // Create and insert the integer alloca.
309 const Type *NewTy = IntegerType::get(AllocaSize*8);
310 NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
311 ConvertUsesToScalar(AI, NewAI, 0);
314 AI->eraseFromParent();
320 // Otherwise, couldn't process this alloca.
326 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
327 /// predicate, do SROA now.
328 void SROA::DoScalarReplacement(AllocationInst *AI,
329 std::vector<AllocationInst*> &WorkList) {
330 DOUT << "Found inst to SROA: " << *AI;
331 SmallVector<AllocaInst*, 32> ElementAllocas;
332 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
333 ElementAllocas.reserve(ST->getNumContainedTypes());
334 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
335 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
337 AI->getName() + "." + utostr(i), AI);
338 ElementAllocas.push_back(NA);
339 WorkList.push_back(NA); // Add to worklist for recursive processing
342 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
343 ElementAllocas.reserve(AT->getNumElements());
344 const Type *ElTy = AT->getElementType();
345 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
346 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
347 AI->getName() + "." + utostr(i), AI);
348 ElementAllocas.push_back(NA);
349 WorkList.push_back(NA); // Add to worklist for recursive processing
353 // Now that we have created the alloca instructions that we want to use,
354 // expand the getelementptr instructions to use them.
356 while (!AI->use_empty()) {
357 Instruction *User = cast<Instruction>(AI->use_back());
358 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) {
359 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas);
360 BCInst->eraseFromParent();
365 // %res = load { i32, i32 }* %alloc
367 // %load.0 = load i32* %alloc.0
368 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
369 // %load.1 = load i32* %alloc.1
370 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
371 // (Also works for arrays instead of structs)
372 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
373 Value *Insert = UndefValue::get(LI->getType());
374 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) {
375 Value *Load = new LoadInst(ElementAllocas[i], "load", LI);
376 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI);
378 LI->replaceAllUsesWith(Insert);
379 LI->eraseFromParent();
384 // store { i32, i32 } %val, { i32, i32 }* %alloc
386 // %val.0 = extractvalue { i32, i32 } %val, 0
387 // store i32 %val.0, i32* %alloc.0
388 // %val.1 = extractvalue { i32, i32 } %val, 1
389 // store i32 %val.1, i32* %alloc.1
390 // (Also works for arrays instead of structs)
391 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
392 Value *Val = SI->getOperand(0);
393 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) {
394 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI);
395 new StoreInst(Extract, ElementAllocas[i], SI);
397 SI->eraseFromParent();
401 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
402 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst>
404 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
406 assert(Idx < ElementAllocas.size() && "Index out of range?");
407 AllocaInst *AllocaToUse = ElementAllocas[Idx];
410 if (GEPI->getNumOperands() == 3) {
411 // Do not insert a new getelementptr instruction with zero indices, only
412 // to have it optimized out later.
413 RepValue = AllocaToUse;
415 // We are indexing deeply into the structure, so we still need a
416 // getelement ptr instruction to finish the indexing. This may be
417 // expanded itself once the worklist is rerun.
419 SmallVector<Value*, 8> NewArgs;
420 NewArgs.push_back(Constant::getNullValue(Type::Int32Ty));
421 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end());
422 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(),
423 NewArgs.end(), "", GEPI);
424 RepValue->takeName(GEPI);
427 // If this GEP is to the start of the aggregate, check for memcpys.
428 if (Idx == 0 && GEPI->hasAllZeroIndices())
429 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas);
431 // Move all of the users over to the new GEP.
432 GEPI->replaceAllUsesWith(RepValue);
433 // Delete the old GEP
434 GEPI->eraseFromParent();
437 // Finally, delete the Alloca instruction
438 AI->eraseFromParent();
443 /// isSafeElementUse - Check to see if this use is an allowed use for a
444 /// getelementptr instruction of an array aggregate allocation. isFirstElt
445 /// indicates whether Ptr is known to the start of the aggregate.
447 void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI,
449 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
451 Instruction *User = cast<Instruction>(*I);
452 switch (User->getOpcode()) {
453 case Instruction::Load: break;
454 case Instruction::Store:
455 // Store is ok if storing INTO the pointer, not storing the pointer
456 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info);
458 case Instruction::GetElementPtr: {
459 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User);
460 bool AreAllZeroIndices = isFirstElt;
461 if (GEP->getNumOperands() > 1) {
462 if (!isa<ConstantInt>(GEP->getOperand(1)) ||
463 !cast<ConstantInt>(GEP->getOperand(1))->isZero())
464 // Using pointer arithmetic to navigate the array.
465 return MarkUnsafe(Info);
467 if (AreAllZeroIndices)
468 AreAllZeroIndices = GEP->hasAllZeroIndices();
470 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info);
471 if (Info.isUnsafe) return;
474 case Instruction::BitCast:
476 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info);
477 if (Info.isUnsafe) return;
480 DOUT << " Transformation preventing inst: " << *User;
481 return MarkUnsafe(Info);
482 case Instruction::Call:
483 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
485 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info);
486 if (Info.isUnsafe) return;
490 DOUT << " Transformation preventing inst: " << *User;
491 return MarkUnsafe(Info);
493 DOUT << " Transformation preventing inst: " << *User;
494 return MarkUnsafe(Info);
497 return; // All users look ok :)
500 /// AllUsersAreLoads - Return true if all users of this value are loads.
501 static bool AllUsersAreLoads(Value *Ptr) {
502 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
504 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load)
509 /// isSafeUseOfAllocation - Check to see if this user is an allowed use for an
510 /// aggregate allocation.
512 void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI,
514 if (BitCastInst *C = dyn_cast<BitCastInst>(User))
515 return isSafeUseOfBitCastedAllocation(C, AI, Info);
517 if (LoadInst *LI = dyn_cast<LoadInst>(User))
518 if (!LI->isVolatile())
519 return;// Loads (returning a first class aggregrate) are always rewritable
521 if (StoreInst *SI = dyn_cast<StoreInst>(User))
522 if (!SI->isVolatile() && SI->getOperand(0) != AI)
523 return;// Store is ok if storing INTO the pointer, not storing the pointer
525 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User);
527 return MarkUnsafe(Info);
529 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI);
531 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>".
533 I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) {
534 return MarkUnsafe(Info);
538 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices??
540 bool IsAllZeroIndices = true;
542 // If the first index is a non-constant index into an array, see if we can
543 // handle it as a special case.
544 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
545 if (!isa<ConstantInt>(I.getOperand())) {
546 IsAllZeroIndices = 0;
547 uint64_t NumElements = AT->getNumElements();
549 // If this is an array index and the index is not constant, we cannot
550 // promote... that is unless the array has exactly one or two elements in
551 // it, in which case we CAN promote it, but we have to canonicalize this
552 // out if this is the only problem.
553 if ((NumElements == 1 || NumElements == 2) &&
554 AllUsersAreLoads(GEPI)) {
555 Info.needsCleanup = true;
556 return; // Canonicalization required!
558 return MarkUnsafe(Info);
562 // Walk through the GEP type indices, checking the types that this indexes
564 for (; I != E; ++I) {
565 // Ignore struct elements, no extra checking needed for these.
566 if (isa<StructType>(*I))
569 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand());
570 if (!IdxVal) return MarkUnsafe(Info);
572 // Are all indices still zero?
573 IsAllZeroIndices &= IdxVal->isZero();
575 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
576 // This GEP indexes an array. Verify that this is an in-range constant
577 // integer. Specifically, consider A[0][i]. We cannot know that the user
578 // isn't doing invalid things like allowing i to index an out-of-range
579 // subscript that accesses A[1]. Because of this, we have to reject SROA
580 // of any accesses into structs where any of the components are variables.
581 if (IdxVal->getZExtValue() >= AT->getNumElements())
582 return MarkUnsafe(Info);
583 } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) {
584 if (IdxVal->getZExtValue() >= VT->getNumElements())
585 return MarkUnsafe(Info);
589 // If there are any non-simple uses of this getelementptr, make sure to reject
591 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info);
594 /// isSafeMemIntrinsicOnAllocation - Return true if the specified memory
595 /// intrinsic can be promoted by SROA. At this point, we know that the operand
596 /// of the memintrinsic is a pointer to the beginning of the allocation.
597 void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI,
598 unsigned OpNo, AllocaInfo &Info) {
599 // If not constant length, give up.
600 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
601 if (!Length) return MarkUnsafe(Info);
603 // If not the whole aggregate, give up.
604 if (Length->getZExtValue() !=
605 TD->getTypePaddedSize(AI->getType()->getElementType()))
606 return MarkUnsafe(Info);
608 // We only know about memcpy/memset/memmove.
609 if (!isa<MemCpyInst>(MI) && !isa<MemSetInst>(MI) && !isa<MemMoveInst>(MI))
610 return MarkUnsafe(Info);
612 // Otherwise, we can transform it. Determine whether this is a memcpy/set
613 // into or out of the aggregate.
615 Info.isMemCpyDst = true;
618 Info.isMemCpySrc = true;
622 /// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast
624 void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI,
626 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end();
628 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) {
629 isSafeUseOfBitCastedAllocation(BCU, AI, Info);
630 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) {
631 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info);
632 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
633 if (SI->isVolatile())
634 return MarkUnsafe(Info);
636 // If storing the entire alloca in one chunk through a bitcasted pointer
637 // to integer, we can transform it. This happens (for example) when you
638 // cast a {i32,i32}* to i64* and store through it. This is similar to the
639 // memcpy case and occurs in various "byval" cases and emulated memcpys.
640 if (isa<IntegerType>(SI->getOperand(0)->getType()) &&
641 TD->getTypePaddedSize(SI->getOperand(0)->getType()) ==
642 TD->getTypePaddedSize(AI->getType()->getElementType())) {
643 Info.isMemCpyDst = true;
646 return MarkUnsafe(Info);
647 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) {
648 if (LI->isVolatile())
649 return MarkUnsafe(Info);
651 // If loading the entire alloca in one chunk through a bitcasted pointer
652 // to integer, we can transform it. This happens (for example) when you
653 // cast a {i32,i32}* to i64* and load through it. This is similar to the
654 // memcpy case and occurs in various "byval" cases and emulated memcpys.
655 if (isa<IntegerType>(LI->getType()) &&
656 TD->getTypePaddedSize(LI->getType()) ==
657 TD->getTypePaddedSize(AI->getType()->getElementType())) {
658 Info.isMemCpySrc = true;
661 return MarkUnsafe(Info);
662 } else if (isa<DbgInfoIntrinsic>(UI)) {
663 // If one user is DbgInfoIntrinsic then check if all users are
664 // DbgInfoIntrinsics.
665 if (OnlyUsedByDbgInfoIntrinsics(BC)) {
666 Info.needsCleanup = true;
673 return MarkUnsafe(Info);
675 if (Info.isUnsafe) return;
679 /// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes
680 /// to its first element. Transform users of the cast to use the new values
682 void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI,
683 SmallVector<AllocaInst*, 32> &NewElts) {
684 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end();
686 Instruction *User = cast<Instruction>(*UI++);
687 if (BitCastInst *BCU = dyn_cast<BitCastInst>(User)) {
688 RewriteBitCastUserOfAlloca(BCU, AI, NewElts);
689 if (BCU->use_empty()) BCU->eraseFromParent();
693 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
694 // This must be memcpy/memmove/memset of the entire aggregate.
695 // Split into one per element.
696 RewriteMemIntrinUserOfAlloca(MI, BCInst, AI, NewElts);
700 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
701 // If this is a store of the entire alloca from an integer, rewrite it.
702 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
706 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
707 // If this is a load of the entire alloca to an integer, rewrite it.
708 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
712 // Otherwise it must be some other user of a gep of the first pointer. Just
713 // leave these alone.
718 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
719 /// Rewrite it to copy or set the elements of the scalarized memory.
720 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst,
722 SmallVector<AllocaInst*, 32> &NewElts) {
724 // If this is a memcpy/memmove, construct the other pointer as the
727 if (MemCpyInst *MCI = dyn_cast<MemCpyInst>(MI)) {
728 if (BCInst == MCI->getRawDest())
729 OtherPtr = MCI->getRawSource();
731 assert(BCInst == MCI->getRawSource());
732 OtherPtr = MCI->getRawDest();
734 } else if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
735 if (BCInst == MMI->getRawDest())
736 OtherPtr = MMI->getRawSource();
738 assert(BCInst == MMI->getRawSource());
739 OtherPtr = MMI->getRawDest();
743 // If there is an other pointer, we want to convert it to the same pointer
744 // type as AI has, so we can GEP through it safely.
746 // It is likely that OtherPtr is a bitcast, if so, remove it.
747 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr))
748 OtherPtr = BC->getOperand(0);
749 // All zero GEPs are effectively bitcasts.
750 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr))
751 if (GEP->hasAllZeroIndices())
752 OtherPtr = GEP->getOperand(0);
754 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr))
755 if (BCE->getOpcode() == Instruction::BitCast)
756 OtherPtr = BCE->getOperand(0);
758 // If the pointer is not the right type, insert a bitcast to the right
760 if (OtherPtr->getType() != AI->getType())
761 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(),
765 // Process each element of the aggregate.
766 Value *TheFn = MI->getOperand(0);
767 const Type *BytePtrTy = MI->getRawDest()->getType();
768 bool SROADest = MI->getRawDest() == BCInst;
770 Constant *Zero = Constant::getNullValue(Type::Int32Ty);
772 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
773 // If this is a memcpy/memmove, emit a GEP of the other element address.
776 Value *Idx[2] = { Zero, ConstantInt::get(Type::Int32Ty, i) };
777 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2,
778 OtherPtr->getNameStr()+"."+utostr(i),
782 Value *EltPtr = NewElts[i];
783 const Type *EltTy =cast<PointerType>(EltPtr->getType())->getElementType();
785 // If we got down to a scalar, insert a load or store as appropriate.
786 if (EltTy->isSingleValueType()) {
787 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
788 Value *Elt = new LoadInst(SROADest ? OtherElt : EltPtr, "tmp",
790 new StoreInst(Elt, SROADest ? EltPtr : OtherElt, MI);
793 assert(isa<MemSetInst>(MI));
795 // If the stored element is zero (common case), just store a null
798 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) {
800 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
802 // If EltTy is a vector type, get the element type.
803 const Type *ValTy = EltTy;
804 if (const VectorType *VTy = dyn_cast<VectorType>(ValTy))
805 ValTy = VTy->getElementType();
807 // Construct an integer with the right value.
808 unsigned EltSize = TD->getTypeSizeInBits(ValTy);
809 APInt OneVal(EltSize, CI->getZExtValue());
810 APInt TotalVal(OneVal);
812 for (unsigned i = 0; 8*i < EltSize; ++i) {
813 TotalVal = TotalVal.shl(8);
817 // Convert the integer value to the appropriate type.
818 StoreVal = ConstantInt::get(TotalVal);
819 if (isa<PointerType>(ValTy))
820 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
821 else if (ValTy->isFloatingPoint())
822 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
823 assert(StoreVal->getType() == ValTy && "Type mismatch!");
825 // If the requested value was a vector constant, create it.
826 if (EltTy != ValTy) {
827 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements();
828 SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
829 StoreVal = ConstantVector::get(&Elts[0], NumElts);
832 new StoreInst(StoreVal, EltPtr, MI);
835 // Otherwise, if we're storing a byte variable, use a memset call for
839 // Cast the element pointer to BytePtrTy.
840 if (EltPtr->getType() != BytePtrTy)
841 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI);
843 // Cast the other pointer (if we have one) to BytePtrTy.
844 if (OtherElt && OtherElt->getType() != BytePtrTy)
845 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(),
848 unsigned EltSize = TD->getTypePaddedSize(EltTy);
850 // Finally, insert the meminst for this element.
851 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
853 SROADest ? EltPtr : OtherElt, // Dest ptr
854 SROADest ? OtherElt : EltPtr, // Src ptr
855 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
858 CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
860 assert(isa<MemSetInst>(MI));
862 EltPtr, MI->getOperand(2), // Dest, Value,
863 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
866 CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
869 MI->eraseFromParent();
872 /// RewriteStoreUserOfWholeAlloca - We found an store of an integer that
873 /// overwrites the entire allocation. Extract out the pieces of the stored
874 /// integer and store them individually.
875 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI,
877 SmallVector<AllocaInst*, 32> &NewElts){
878 // Extract each element out of the integer according to its structure offset
879 // and store the element value to the individual alloca.
880 Value *SrcVal = SI->getOperand(0);
881 const Type *AllocaEltTy = AI->getType()->getElementType();
882 uint64_t AllocaSizeBits = TD->getTypePaddedSizeInBits(AllocaEltTy);
884 // If this isn't a store of an integer to the whole alloca, it may be a store
885 // to the first element. Just ignore the store in this case and normal SROA
887 if (!isa<IntegerType>(SrcVal->getType()) ||
888 TD->getTypePaddedSizeInBits(SrcVal->getType()) != AllocaSizeBits)
891 DOUT << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << *SI;
893 // There are two forms here: AI could be an array or struct. Both cases
894 // have different ways to compute the element offset.
895 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
896 const StructLayout *Layout = TD->getStructLayout(EltSTy);
898 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
899 // Get the number of bits to shift SrcVal to get the value.
900 const Type *FieldTy = EltSTy->getElementType(i);
901 uint64_t Shift = Layout->getElementOffsetInBits(i);
903 if (TD->isBigEndian())
904 Shift = AllocaSizeBits-Shift-TD->getTypePaddedSizeInBits(FieldTy);
906 Value *EltVal = SrcVal;
908 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
909 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
910 "sroa.store.elt", SI);
913 // Truncate down to an integer of the right size.
914 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
916 // Ignore zero sized fields like {}, they obviously contain no data.
917 if (FieldSizeBits == 0) continue;
919 if (FieldSizeBits != AllocaSizeBits)
920 EltVal = new TruncInst(EltVal, IntegerType::get(FieldSizeBits), "", SI);
921 Value *DestField = NewElts[i];
922 if (EltVal->getType() == FieldTy) {
923 // Storing to an integer field of this size, just do it.
924 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) {
925 // Bitcast to the right element type (for fp/vector values).
926 EltVal = new BitCastInst(EltVal, FieldTy, "", SI);
928 // Otherwise, bitcast the dest pointer (for aggregates).
929 DestField = new BitCastInst(DestField,
930 PointerType::getUnqual(EltVal->getType()),
933 new StoreInst(EltVal, DestField, SI);
937 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
938 const Type *ArrayEltTy = ATy->getElementType();
939 uint64_t ElementOffset = TD->getTypePaddedSizeInBits(ArrayEltTy);
940 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
944 if (TD->isBigEndian())
945 Shift = AllocaSizeBits-ElementOffset;
949 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
950 // Ignore zero sized fields like {}, they obviously contain no data.
951 if (ElementSizeBits == 0) continue;
953 Value *EltVal = SrcVal;
955 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
956 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
957 "sroa.store.elt", SI);
960 // Truncate down to an integer of the right size.
961 if (ElementSizeBits != AllocaSizeBits)
962 EltVal = new TruncInst(EltVal, IntegerType::get(ElementSizeBits),"",SI);
963 Value *DestField = NewElts[i];
964 if (EltVal->getType() == ArrayEltTy) {
965 // Storing to an integer field of this size, just do it.
966 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) {
967 // Bitcast to the right element type (for fp/vector values).
968 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI);
970 // Otherwise, bitcast the dest pointer (for aggregates).
971 DestField = new BitCastInst(DestField,
972 PointerType::getUnqual(EltVal->getType()),
975 new StoreInst(EltVal, DestField, SI);
977 if (TD->isBigEndian())
978 Shift -= ElementOffset;
980 Shift += ElementOffset;
984 SI->eraseFromParent();
987 /// RewriteLoadUserOfWholeAlloca - We found an load of the entire allocation to
988 /// an integer. Load the individual pieces to form the aggregate value.
989 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI,
990 SmallVector<AllocaInst*, 32> &NewElts) {
991 // Extract each element out of the NewElts according to its structure offset
992 // and form the result value.
993 const Type *AllocaEltTy = AI->getType()->getElementType();
994 uint64_t AllocaSizeBits = TD->getTypePaddedSizeInBits(AllocaEltTy);
996 // If this isn't a load of the whole alloca to an integer, it may be a load
997 // of the first element. Just ignore the load in this case and normal SROA
999 if (!isa<IntegerType>(LI->getType()) ||
1000 TD->getTypePaddedSizeInBits(LI->getType()) != AllocaSizeBits)
1003 DOUT << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << *LI;
1005 // There are two forms here: AI could be an array or struct. Both cases
1006 // have different ways to compute the element offset.
1007 const StructLayout *Layout = 0;
1008 uint64_t ArrayEltBitOffset = 0;
1009 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
1010 Layout = TD->getStructLayout(EltSTy);
1012 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
1013 ArrayEltBitOffset = TD->getTypePaddedSizeInBits(ArrayEltTy);
1016 Value *ResultVal = Constant::getNullValue(LI->getType());
1018 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1019 // Load the value from the alloca. If the NewElt is an aggregate, cast
1020 // the pointer to an integer of the same size before doing the load.
1021 Value *SrcField = NewElts[i];
1022 const Type *FieldTy =
1023 cast<PointerType>(SrcField->getType())->getElementType();
1024 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
1026 // Ignore zero sized fields like {}, they obviously contain no data.
1027 if (FieldSizeBits == 0) continue;
1029 const IntegerType *FieldIntTy = IntegerType::get(FieldSizeBits);
1030 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() &&
1031 !isa<VectorType>(FieldTy))
1032 SrcField = new BitCastInst(SrcField, PointerType::getUnqual(FieldIntTy),
1034 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
1036 // If SrcField is a fp or vector of the right size but that isn't an
1037 // integer type, bitcast to an integer so we can shift it.
1038 if (SrcField->getType() != FieldIntTy)
1039 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
1041 // Zero extend the field to be the same size as the final alloca so that
1042 // we can shift and insert it.
1043 if (SrcField->getType() != ResultVal->getType())
1044 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
1046 // Determine the number of bits to shift SrcField.
1048 if (Layout) // Struct case.
1049 Shift = Layout->getElementOffsetInBits(i);
1051 Shift = i*ArrayEltBitOffset;
1053 if (TD->isBigEndian())
1054 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
1057 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
1058 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
1061 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
1064 LI->replaceAllUsesWith(ResultVal);
1065 LI->eraseFromParent();
1069 /// HasPadding - Return true if the specified type has any structure or
1070 /// alignment padding, false otherwise.
1071 static bool HasPadding(const Type *Ty, const TargetData &TD) {
1072 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1073 const StructLayout *SL = TD.getStructLayout(STy);
1074 unsigned PrevFieldBitOffset = 0;
1075 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1076 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
1078 // Padding in sub-elements?
1079 if (HasPadding(STy->getElementType(i), TD))
1082 // Check to see if there is any padding between this element and the
1085 unsigned PrevFieldEnd =
1086 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
1087 if (PrevFieldEnd < FieldBitOffset)
1091 PrevFieldBitOffset = FieldBitOffset;
1094 // Check for tail padding.
1095 if (unsigned EltCount = STy->getNumElements()) {
1096 unsigned PrevFieldEnd = PrevFieldBitOffset +
1097 TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
1098 if (PrevFieldEnd < SL->getSizeInBits())
1102 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1103 return HasPadding(ATy->getElementType(), TD);
1104 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1105 return HasPadding(VTy->getElementType(), TD);
1107 return TD.getTypeSizeInBits(Ty) != TD.getTypePaddedSizeInBits(Ty);
1110 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
1111 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
1112 /// or 1 if safe after canonicalization has been performed.
1114 int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) {
1115 // Loop over the use list of the alloca. We can only transform it if all of
1116 // the users are safe to transform.
1119 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end();
1121 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info);
1122 if (Info.isUnsafe) {
1123 DOUT << "Cannot transform: " << *AI << " due to user: " << **I;
1128 // Okay, we know all the users are promotable. If the aggregate is a memcpy
1129 // source and destination, we have to be careful. In particular, the memcpy
1130 // could be moving around elements that live in structure padding of the LLVM
1131 // types, but may actually be used. In these cases, we refuse to promote the
1133 if (Info.isMemCpySrc && Info.isMemCpyDst &&
1134 HasPadding(AI->getType()->getElementType(), *TD))
1137 // If we require cleanup, return 1, otherwise return 3.
1138 return Info.needsCleanup ? 1 : 3;
1141 /// CleanupGEP - GEP is used by an Alloca, which can be prompted after the GEP
1142 /// is canonicalized here.
1143 void SROA::CleanupGEP(GetElementPtrInst *GEPI) {
1144 gep_type_iterator I = gep_type_begin(GEPI);
1147 const ArrayType *AT = dyn_cast<ArrayType>(*I);
1151 uint64_t NumElements = AT->getNumElements();
1153 if (isa<ConstantInt>(I.getOperand()))
1156 if (NumElements == 1) {
1157 GEPI->setOperand(2, Constant::getNullValue(Type::Int32Ty));
1161 assert(NumElements == 2 && "Unhandled case!");
1162 // All users of the GEP must be loads. At each use of the GEP, insert
1163 // two loads of the appropriate indexed GEP and select between them.
1164 Value *IsOne = new ICmpInst(ICmpInst::ICMP_NE, I.getOperand(),
1165 Constant::getNullValue(I.getOperand()->getType()),
1167 // Insert the new GEP instructions, which are properly indexed.
1168 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end());
1169 Indices[1] = Constant::getNullValue(Type::Int32Ty);
1170 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
1173 GEPI->getName()+".0", GEPI);
1174 Indices[1] = ConstantInt::get(Type::Int32Ty, 1);
1175 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
1178 GEPI->getName()+".1", GEPI);
1179 // Replace all loads of the variable index GEP with loads from both
1180 // indexes and a select.
1181 while (!GEPI->use_empty()) {
1182 LoadInst *LI = cast<LoadInst>(GEPI->use_back());
1183 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI);
1184 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI);
1185 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI);
1186 LI->replaceAllUsesWith(R);
1187 LI->eraseFromParent();
1189 GEPI->eraseFromParent();
1193 /// CleanupAllocaUsers - If SROA reported that it can promote the specified
1194 /// allocation, but only if cleaned up, perform the cleanups required.
1195 void SROA::CleanupAllocaUsers(AllocationInst *AI) {
1196 // At this point, we know that the end result will be SROA'd and promoted, so
1197 // we can insert ugly code if required so long as sroa+mem2reg will clean it
1199 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
1202 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U))
1204 else if (Instruction *I = dyn_cast<Instruction>(U)) {
1205 SmallVector<DbgInfoIntrinsic *, 2> DbgInUses;
1206 if (OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) {
1207 // Safe to remove debug info uses.
1208 while (!DbgInUses.empty()) {
1209 DbgInfoIntrinsic *DI = DbgInUses.back(); DbgInUses.pop_back();
1210 DI->eraseFromParent();
1212 I->eraseFromParent();
1218 /// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at
1219 /// the offset specified by Offset (which is specified in bytes).
1221 /// There are two cases we handle here:
1222 /// 1) A union of vector types of the same size and potentially its elements.
1223 /// Here we turn element accesses into insert/extract element operations.
1224 /// This promotes a <4 x float> with a store of float to the third element
1225 /// into a <4 x float> that uses insert element.
1226 /// 2) A fully general blob of memory, which we turn into some (potentially
1227 /// large) integer type with extract and insert operations where the loads
1228 /// and stores would mutate the memory.
1229 static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy,
1230 unsigned AllocaSize, const TargetData &TD) {
1231 // If this could be contributing to a vector, analyze it.
1232 if (VecTy != Type::VoidTy) { // either null or a vector type.
1234 // If the In type is a vector that is the same size as the alloca, see if it
1235 // matches the existing VecTy.
1236 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) {
1237 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) {
1238 // If we're storing/loading a vector of the right size, allow it as a
1239 // vector. If this the first vector we see, remember the type so that
1240 // we know the element size.
1245 } else if (In == Type::FloatTy || In == Type::DoubleTy ||
1246 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 &&
1247 isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
1248 // If we're accessing something that could be an element of a vector, see
1249 // if the implied vector agrees with what we already have and if Offset is
1250 // compatible with it.
1251 unsigned EltSize = In->getPrimitiveSizeInBits()/8;
1252 if (Offset % EltSize == 0 &&
1253 AllocaSize % EltSize == 0 &&
1255 cast<VectorType>(VecTy)->getElementType()
1256 ->getPrimitiveSizeInBits()/8 == EltSize)) {
1258 VecTy = VectorType::get(In, AllocaSize/EltSize);
1264 // Otherwise, we have a case that we can't handle with an optimized vector
1265 // form. We can still turn this into a large integer.
1266 VecTy = Type::VoidTy;
1269 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
1270 /// its accesses to use a to single vector type, return true, and set VecTy to
1271 /// the new type. If we could convert the alloca into a single promotable
1272 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a
1273 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
1274 /// is the current offset from the base of the alloca being analyzed.
1276 /// If we see at least one access to the value that is as a vector type, set the
1279 bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy,
1280 bool &SawVec, uint64_t Offset,
1281 unsigned AllocaSize) {
1282 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1283 Instruction *User = cast<Instruction>(*UI);
1285 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1286 // Don't break volatile loads.
1287 if (LI->isVolatile())
1289 MergeInType(LI->getType(), Offset, VecTy, AllocaSize, *TD);
1290 SawVec |= isa<VectorType>(LI->getType());
1294 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1295 // Storing the pointer, not into the value?
1296 if (SI->getOperand(0) == V || SI->isVolatile()) return 0;
1297 MergeInType(SI->getOperand(0)->getType(), Offset, VecTy, AllocaSize, *TD);
1298 SawVec |= isa<VectorType>(SI->getOperand(0)->getType());
1302 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
1303 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset,
1306 IsNotTrivial = true;
1310 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1311 // If this is a GEP with a variable indices, we can't handle it.
1312 if (!GEP->hasAllConstantIndices())
1315 // Compute the offset that this GEP adds to the pointer.
1316 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
1317 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(),
1318 &Indices[0], Indices.size());
1319 // See if all uses can be converted.
1320 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset,
1323 IsNotTrivial = true;
1327 // If this is a constant sized memset of a constant value (e.g. 0) we can
1329 if (isa<MemSetInst>(User) &&
1330 // Store of constant value.
1331 isa<ConstantInt>(User->getOperand(2)) &&
1332 // Store with constant size.
1333 isa<ConstantInt>(User->getOperand(3))) {
1334 VecTy = Type::VoidTy;
1335 IsNotTrivial = true;
1339 // Otherwise, we cannot handle this!
1347 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
1348 /// directly. This happens when we are converting an "integer union" to a
1349 /// single integer scalar, or when we are converting a "vector union" to a
1350 /// vector with insert/extractelement instructions.
1352 /// Offset is an offset from the original alloca, in bits that need to be
1353 /// shifted to the right. By the end of this, there should be no uses of Ptr.
1354 void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) {
1355 while (!Ptr->use_empty()) {
1356 Instruction *User = cast<Instruction>(Ptr->use_back());
1358 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
1359 ConvertUsesToScalar(CI, NewAI, Offset);
1360 CI->eraseFromParent();
1364 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1365 // Compute the offset that this GEP adds to the pointer.
1366 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
1367 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(),
1368 &Indices[0], Indices.size());
1369 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
1370 GEP->eraseFromParent();
1374 IRBuilder<> Builder(User->getParent(), User);
1376 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1377 // The load is a bit extract from NewAI shifted right by Offset bits.
1378 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp");
1380 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
1381 LI->replaceAllUsesWith(NewLoadVal);
1382 LI->eraseFromParent();
1386 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1387 assert(SI->getOperand(0) != Ptr && "Consistency error!");
1388 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").c_str());
1389 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
1391 Builder.CreateStore(New, NewAI);
1392 SI->eraseFromParent();
1396 // If this is a constant sized memset of a constant value (e.g. 0) we can
1397 // transform it into a store of the expanded constant value.
1398 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
1399 assert(MSI->getRawDest() == Ptr && "Consistency error!");
1400 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
1401 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
1403 // Compute the value replicated the right number of times.
1404 APInt APVal(NumBytes*8, Val);
1406 // Splat the value if non-zero.
1408 for (unsigned i = 1; i != NumBytes; ++i)
1409 APVal |= APVal << 8;
1411 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").c_str());
1412 Value *New = ConvertScalar_InsertValue(ConstantInt::get(APVal), Old,
1414 Builder.CreateStore(New, NewAI);
1415 MSI->eraseFromParent();
1420 assert(0 && "Unsupported operation!");
1425 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
1426 /// or vector value FromVal, extracting the bits from the offset specified by
1427 /// Offset. This returns the value, which is of type ToType.
1429 /// This happens when we are converting an "integer union" to a single
1430 /// integer scalar, or when we are converting a "vector union" to a vector with
1431 /// insert/extractelement instructions.
1433 /// Offset is an offset from the original alloca, in bits that need to be
1434 /// shifted to the right.
1435 Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType,
1436 uint64_t Offset, IRBuilder<> &Builder) {
1437 // If the load is of the whole new alloca, no conversion is needed.
1438 if (FromVal->getType() == ToType && Offset == 0)
1441 // If the result alloca is a vector type, this is either an element
1442 // access or a bitcast to another vector type of the same size.
1443 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) {
1444 if (isa<VectorType>(ToType))
1445 return Builder.CreateBitCast(FromVal, ToType, "tmp");
1447 // Otherwise it must be an element access.
1450 unsigned EltSize = TD->getTypePaddedSizeInBits(VTy->getElementType());
1451 Elt = Offset/EltSize;
1452 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
1454 // Return the element extracted out of it.
1455 Value *V = Builder.CreateExtractElement(FromVal,
1456 ConstantInt::get(Type::Int32Ty,Elt),
1458 if (V->getType() != ToType)
1459 V = Builder.CreateBitCast(V, ToType, "tmp");
1463 // If ToType is a first class aggregate, extract out each of the pieces and
1464 // use insertvalue's to form the FCA.
1465 if (const StructType *ST = dyn_cast<StructType>(ToType)) {
1466 const StructLayout &Layout = *TD->getStructLayout(ST);
1467 Value *Res = UndefValue::get(ST);
1468 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
1469 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
1470 Offset+Layout.getElementOffsetInBits(i),
1472 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
1477 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
1478 uint64_t EltSize = TD->getTypePaddedSizeInBits(AT->getElementType());
1479 Value *Res = UndefValue::get(AT);
1480 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1481 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
1482 Offset+i*EltSize, Builder);
1483 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
1488 // Otherwise, this must be a union that was converted to an integer value.
1489 const IntegerType *NTy = cast<IntegerType>(FromVal->getType());
1491 // If this is a big-endian system and the load is narrower than the
1492 // full alloca type, we need to do a shift to get the right bits.
1494 if (TD->isBigEndian()) {
1495 // On big-endian machines, the lowest bit is stored at the bit offset
1496 // from the pointer given by getTypeStoreSizeInBits. This matters for
1497 // integers with a bitwidth that is not a multiple of 8.
1498 ShAmt = TD->getTypeStoreSizeInBits(NTy) -
1499 TD->getTypeStoreSizeInBits(ToType) - Offset;
1504 // Note: we support negative bitwidths (with shl) which are not defined.
1505 // We do this to support (f.e.) loads off the end of a structure where
1506 // only some bits are used.
1507 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
1508 FromVal = Builder.CreateLShr(FromVal, ConstantInt::get(FromVal->getType(),
1510 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
1511 FromVal = Builder.CreateShl(FromVal, ConstantInt::get(FromVal->getType(),
1514 // Finally, unconditionally truncate the integer to the right width.
1515 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType);
1516 if (LIBitWidth < NTy->getBitWidth())
1517 FromVal = Builder.CreateTrunc(FromVal, IntegerType::get(LIBitWidth), "tmp");
1518 else if (LIBitWidth > NTy->getBitWidth())
1519 FromVal = Builder.CreateZExt(FromVal, IntegerType::get(LIBitWidth), "tmp");
1521 // If the result is an integer, this is a trunc or bitcast.
1522 if (isa<IntegerType>(ToType)) {
1524 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) {
1525 // Just do a bitcast, we know the sizes match up.
1526 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp");
1528 // Otherwise must be a pointer.
1529 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp");
1531 assert(FromVal->getType() == ToType && "Didn't convert right?");
1536 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
1537 /// or vector value "Old" at the offset specified by Offset.
1539 /// This happens when we are converting an "integer union" to a
1540 /// single integer scalar, or when we are converting a "vector union" to a
1541 /// vector with insert/extractelement instructions.
1543 /// Offset is an offset from the original alloca, in bits that need to be
1544 /// shifted to the right.
1545 Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old,
1546 uint64_t Offset, IRBuilder<> &Builder) {
1548 // Convert the stored type to the actual type, shift it left to insert
1549 // then 'or' into place.
1550 const Type *AllocaType = Old->getType();
1552 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
1553 // If the result alloca is a vector type, this is either an element
1554 // access or a bitcast to another vector type.
1555 if (isa<VectorType>(SV->getType())) {
1556 SV = Builder.CreateBitCast(SV, AllocaType, "tmp");
1558 // Must be an element insertion.
1559 unsigned Elt = Offset/TD->getTypePaddedSizeInBits(VTy->getElementType());
1561 if (SV->getType() != VTy->getElementType())
1562 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp");
1564 SV = Builder.CreateInsertElement(Old, SV,
1565 ConstantInt::get(Type::Int32Ty, Elt),
1571 // If SV is a first-class aggregate value, insert each value recursively.
1572 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) {
1573 const StructLayout &Layout = *TD->getStructLayout(ST);
1574 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
1575 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
1576 Old = ConvertScalar_InsertValue(Elt, Old,
1577 Offset+Layout.getElementOffsetInBits(i),
1583 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
1584 uint64_t EltSize = TD->getTypePaddedSizeInBits(AT->getElementType());
1585 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1586 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
1587 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
1592 // If SV is a float, convert it to the appropriate integer type.
1593 // If it is a pointer, do the same.
1594 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType());
1595 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType);
1596 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType());
1597 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType);
1598 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType()))
1599 SV = Builder.CreateBitCast(SV, IntegerType::get(SrcWidth), "tmp");
1600 else if (isa<PointerType>(SV->getType()))
1601 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(), "tmp");
1603 // Zero extend or truncate the value if needed.
1604 if (SV->getType() != AllocaType) {
1605 if (SV->getType()->getPrimitiveSizeInBits() <
1606 AllocaType->getPrimitiveSizeInBits())
1607 SV = Builder.CreateZExt(SV, AllocaType, "tmp");
1609 // Truncation may be needed if storing more than the alloca can hold
1610 // (undefined behavior).
1611 SV = Builder.CreateTrunc(SV, AllocaType, "tmp");
1612 SrcWidth = DestWidth;
1613 SrcStoreWidth = DestStoreWidth;
1617 // If this is a big-endian system and the store is narrower than the
1618 // full alloca type, we need to do a shift to get the right bits.
1620 if (TD->isBigEndian()) {
1621 // On big-endian machines, the lowest bit is stored at the bit offset
1622 // from the pointer given by getTypeStoreSizeInBits. This matters for
1623 // integers with a bitwidth that is not a multiple of 8.
1624 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
1629 // Note: we support negative bitwidths (with shr) which are not defined.
1630 // We do this to support (f.e.) stores off the end of a structure where
1631 // only some bits in the structure are set.
1632 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
1633 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
1634 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt), "tmp");
1636 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
1637 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt), "tmp");
1638 Mask = Mask.lshr(-ShAmt);
1641 // Mask out the bits we are about to insert from the old value, and or
1643 if (SrcWidth != DestWidth) {
1644 assert(DestWidth > SrcWidth);
1645 Old = Builder.CreateAnd(Old, ConstantInt::get(~Mask), "mask");
1646 SV = Builder.CreateOr(Old, SV, "ins");
1653 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
1654 /// some part of a constant global variable. This intentionally only accepts
1655 /// constant expressions because we don't can't rewrite arbitrary instructions.
1656 static bool PointsToConstantGlobal(Value *V) {
1657 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
1658 return GV->isConstant();
1659 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
1660 if (CE->getOpcode() == Instruction::BitCast ||
1661 CE->getOpcode() == Instruction::GetElementPtr)
1662 return PointsToConstantGlobal(CE->getOperand(0));
1666 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
1667 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
1668 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
1669 /// track of whether it moves the pointer (with isOffset) but otherwise traverse
1670 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
1671 /// the alloca, and if the source pointer is a pointer to a constant global, we
1672 /// can optimize this.
1673 static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy,
1675 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1676 if (LoadInst *LI = dyn_cast<LoadInst>(*UI))
1677 // Ignore non-volatile loads, they are always ok.
1678 if (!LI->isVolatile())
1681 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) {
1682 // If uses of the bitcast are ok, we are ok.
1683 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset))
1687 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
1688 // If the GEP has all zero indices, it doesn't offset the pointer. If it
1689 // doesn't, it does.
1690 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
1691 isOffset || !GEP->hasAllZeroIndices()))
1696 // If this is isn't our memcpy/memmove, reject it as something we can't
1698 if (!isa<MemCpyInst>(*UI) && !isa<MemMoveInst>(*UI))
1701 // If we already have seen a copy, reject the second one.
1702 if (TheCopy) return false;
1704 // If the pointer has been offset from the start of the alloca, we can't
1705 // safely handle this.
1706 if (isOffset) return false;
1708 // If the memintrinsic isn't using the alloca as the dest, reject it.
1709 if (UI.getOperandNo() != 1) return false;
1711 MemIntrinsic *MI = cast<MemIntrinsic>(*UI);
1713 // If the source of the memcpy/move is not a constant global, reject it.
1714 if (!PointsToConstantGlobal(MI->getOperand(2)))
1717 // Otherwise, the transform is safe. Remember the copy instruction.
1723 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
1724 /// modified by a copy from a constant global. If we can prove this, we can
1725 /// replace any uses of the alloca with uses of the global directly.
1726 Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) {
1727 Instruction *TheCopy = 0;
1728 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false))